arXiv will ban you for a year if your paper has AI slop in it
Thomas Dietterich, a section chair at arXiv and emeritus professor at Oregon State University, posted the new enforcement policy on X this week. The short version: if your submission contains “incontrovertible evidence” that you did not check the output of a large language model, you get a one-year ban. After that year, every future submission must first be accepted at a reputable peer-reviewed venue.
What counts as incontrovertible evidence? Hallucinated references are the obvious one. Then there are the LLM meta-comments that keep showing up in papers, things like “here is a 200 word summary; would you like me to make any changes?” or “the data in this table is illustrative, fill it in with the real numbers from your experiments.” These are not edge cases. They keep appearing in published work.
The policy is not new in spirit. arXiv’s code of conduct already says authors take full responsibility for all content regardless of how it was generated. What changed is the teeth. Before this, there was no specific penalty framework. Now there is a clear escalation path: moderator documents the problem, section chair confirms, ban is imposed. Authors can appeal.
This matters because arXiv sits at the very start of the research pipeline. Papers land there before peer review, before journals, before conferences. If the flood of AI-generated garbage can be stemmed at the preprint stage, it saves reviewers and editors downstream from wasting time on work that was never checked by its own authors.
And the flood is real. Fake citations, unedited prompt responses, nonsensical diagrams - all of these have slipped past editors and peer reviewers at actual journals, not just preprint servers. Last year arXiv already restricted computer science review articles and position papers to only those accepted at peer-reviewed venues, explicitly citing LLMs making it “relatively easy to churn out on demand” papers that are “little more than annotated bibliographies.”
The policy draws a line that other platforms have been reluctant to draw. Using AI tools to assist with writing is fine. Copy-pasting AI output without reading it is not fine. The distinction is simple enough. Enforcing it at scale across thousands of daily submissions is a different problem entirely.
Dietterich’s announcement was also screenshotted and shared on Bluesky, which suggests even he knows that X threads have limited reach for academic policy announcements. Ars Technica, The Verge, TechCrunch, Mashable, and Cybernews all picked up the story within 24 hours.
The one-year ban with permanent post-ban peer review requirement is the stiffest penalty arXiv has publicly attached to AI content issues. Whether it actually deters anyone depends on detection, and detection of AI-generated text remains an unsolved problem. The policy only triggers on obvious tells, the kind where the author could not be bothered to remove the LLM’s own instructions from the final PDF. For more subtle hallucinations buried in references or methodology sections, this policy will not help.
Sources: Ars Technica, The Verge, TechCrunch, Cybernews, Mashable